RadiusDirectionPosteriors(Under Refactoring for cleaning up)

Set up

** Virtual Environment Without conda **

git clone https://github.com/ChangYong-Oh/RadiusDirectionPosteriors.git
cd RadiusDirectionPosteriors
source setup_pip.sh

** Virtual Environment With conda **

conda create -n RadiusDirectionPosteriors python=2.7.14 anaconda --yes
cd "`which python | xargs dirname | xargs dirname`/envs/RadiusDirectionPosteriors"
git clone https://github.com/ChangYong-Oh/RadiusDirectionPosteriors.git
source RadiusDirectionPosteriors/setup_conda.sh
Default python should be the anaconda python.

Different python version is possible. For avaialbe version search

conda search "^python$"

** Import in existing Python environment **

Or to be able to import this code in an existing Python environment, go:

pip install -e git+https://github.com/ChangYong-Oh/RadiusDirectionPosteriors.git#egg=RadiusDirectionPosteriors

Directories

Data directory and experiment result directory can be set in

RadiusDirectionPosteriors/utils/dir_utils.py

UCI regression Training

All below files can be called with arguments and information about arguments can be retrieved by calling those file with --help option.

python RadiusDirectionPosteriors/train_double_uci.py --help

For dataset splitting, it is recommended to follow splitting given in

https://github.com/yaringal/DropoutUncertaintyExps

Compression Training

You can train model with compression with file

RadiusDirectionPosteriors/main.py

Arguments

model_type : LeNetFC, LeNet5-Bundle, LeNet5-Flatten
prior_file : json file contraining prior information, when only a filename is given without an absolute path, it looks for a json file in prior_json directory.
epochs : number of epochs
batch_size : default 100
lr : default 0.001
gpu : when this is given, then gpu is used
model_file : only needed when you want to continue training initialized with setting in model_file
eval : Using setting in model_file, model is evaluated.

Example

python RadiusDirectionPosteriors/main.py --model_type LeNet5 --prior_file HalfCauchy-fc0.01-conv0.01.json --epochs 200

For LeNetFC, LeNet5-Bundle, LeNet5-Flatten, cpu is faster.

Compression plotting

In each of following files, you can plot logarithm of mode of radius posteriors. Threshold and model_file can be specified in files.